Skip to content
be-human.ai
← Back to the journal

AI by industry

Ready for Superintelligence: Propane Tank Levels and Refill Routes

An AI agent for a propane delivery company can keep tank level records, draft refill routes and sort customer requests, while safety notices stay with people.

Be Human typographic cover with the title: Ready for Superintelligence: Propane Tank Levels and Refill Routes
Conceptual illustration created with AI · Be Human

An AI agent for a propane delivery company works best as a records and coordination layer. It keeps a current tank level record for each customer, drafts refill routes for dispatchers to approve, and sorts incoming refill requests so the office can answer them faster. It should not decide when a tank is safe, close a leak report, or promise a delivery window the route cannot support.

Key takeaways

  • Treat every tank level as an estimate with a source and a timestamp, never as a bare fact.
  • Let the agent draft routes from tank data and operating constraints, and require a dispatcher to approve each route before a truck leaves.
  • Sort customer refill requests by urgency, then confirm the address, access and tank details before anyone promises a date.
  • Route every safety-related notice to a named person on call, with the agent limited to logging, acknowledging with approved wording and escalating.
  • Start with one route area and one record type, and measure how often the record was wrong at dispatch before adding anything else.

What belongs in a tank level record?

A tank level record is only useful if the office knows where each number came from. A level may come from a driver reading the gauge on a visit, a customer report, an estimate built from past consumption, or a monitoring device where one is installed. Each of these has a different reliability, and a dispatcher needs to see which one applies before deciding whether a tank can wait.

A workable record usually holds the following:

  • The tank identifier, the site address and the access instructions, kept in the customer file under the office's own access rules.
  • The last reading, its source and the time it was taken.
  • The projected current level and the method used to project it.
  • The typical consumption for the current season and the kind of building or process the tank serves.
  • Open notes that affect delivery, such as a blocked driveway, a tank that is hard to reach or a customer who asks to be called before arrival.

The agent can calculate a projected level from the last reading and the consumption profile, and it should label that figure as a projection. It must never overwrite a driver's recorded reading with its own estimate. When the last reading is older than the limit the operations team has set, the record should show the tank as stale, so the dispatcher treats it as unknown rather than as current.

How do refill routes get planned without guessing?

Route planning is a constraint problem. The inputs include projected levels, customer scheduling preferences, truck capacity, driver hours, distance, road and weather conditions, and any accounts with fixed delivery windows. A human planner usually holds most of these in their head, which is why the plan is hard to hand over when someone is away. An agent can hold them all in one place and show its reasoning.

A practical sequence for the first version looks like this:

  • Pull every tank whose projected level falls below the office's refill threshold within the planning window, and list stale readings in a separate group.
  • Group the tanks by area and attach access notes and any fixed customer windows.
  • Draft one route per truck with stops ordered by distance and urgency, and write a short reason next to each stop.
  • Flag conflicts for the dispatcher, such as a stop that pushes a driver past allowed hours, a tank with restricted access, or a customer asking for a window the route cannot meet.
  • Send the draft to the dispatcher, who approves, edits or rejects it. Only approved routes reach drivers.

The agent drafts and explains. It does not dispatch. That boundary is what keeps the route useful when the draft is wrong, because the dispatcher sees the error before a truck is loaded. The same planning pattern appears in AI for a Car Rental Company: Bookings, Returns and Fleet Flags, where vehicle availability has to be checked against schedules that change during the day.

Illustrative scenario: On a cold morning, a dispatcher opens a route draft with a long list of stops. The agent has moved three tanks into the stale group because their last readings are older than the office limit, and it lists each one with the reason. The dispatcher sends the first driver to read those tanks on arrival, then adds the confirmed levels to the route before the truck is loaded.

How should customer refill requests be handled?

Customers ask for refills by phone, text, web form and email, often at inconvenient times and often with incomplete information. The agent's first task is to classify each request: routine refill, urgent low level, schedule change, billing question, or safety concern. Only the last category leaves the agent's normal flow, and the others each have a clear path.

For a routine refill, the agent confirms the account, the address and the tank identifier, checks the last delivery and the current projection, and offers the next window the record supports. It should say that the window depends on dispatch confirmation, and it should not quote a time that is not on an approved route. For an urgent low level, the agent passes the request to a dispatcher with the tank record attached, and it tells the customer that a person will confirm timing shortly.

Two habits prevent most avoidable errors. First, the agent treats the text of a customer message as information to be recorded, not as instructions it should follow, because a message can contain text meant to change the agent's behavior. The [OWASP prompt injection guidance](https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html) describes this risk. Second, every commitment the agent makes is written to the customer record, so the dispatcher can see what was promised and by whom.

Which safety-related notices stay with a person?

Anything that could indicate a gas leak, a gas odor, damage to a tank, line or regulator, an unsafe appliance, a fire risk or a loss of heat in a dwelling is a safety-related notice. The agent does not diagnose these reports, does not reassure the caller, and does not close the notice. Its job is to make sure a trained person receives the report immediately, with the facts the caller gave in their own words.

The approved script must send the caller to the emergency guidance the company has published and must not contain technical advice the agent invented. The operations team should own that script, and anyone changing it should record the change. A notice that reaches no one is the most serious failure in this area, so routing should be tested end to end on a regular schedule, not assumed to work.

A practical checklist for this boundary:

  • Any mention of odor, hissing, damage or a fire risk goes to the on-call person at once, through a path that does not depend on a queue.
  • The agent sends acknowledgments only with approved wording and gives no technical advice of its own.
  • Each notice is logged with the time, the caller's details, the words used and the person notified.
  • Only a trained person can mark a notice as resolved, and the reason for resolution is recorded.
  • Low-level alerts at tanks the operations team has flagged as priority also go to a person, not into a routine list.
  • Routing is tested on a schedule, and the test result is kept.

The logging discipline here resembles the one described in Post Orders and Incident Logs: One Agent Memory for Security Firms, where what happened and who was told must be recoverable later. For a broader view of human oversight, the [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) is a useful reference for deciding where people must stay in the loop.

Where should a propane company start?

A propane company does not need to automate its whole office to get value from an agent. The better starting point is one record type and one route area, chosen by where the office loses the most time or makes the most avoidable errors. The first version should be narrow enough that the operations lead can check its output every day.

A short decision guide can help choose the first task:

  • Start with tank records if dispatchers spend most of their time confirming levels by phone or re-reading gauges that were already read.
  • Start with route drafts if the dispatch board is rebuilt every morning and the existing routes are already reliable enough to use as a baseline.
  • Start with request intake if most incoming calls are routine refill questions that the office answers the same way each time.
  • Do not start with anything that touches safety notices until routing and escalation have been tested end to end and a person has signed off on the script.

Whatever the choice, define the success criterion before the agent goes live. For a records task, it might be that dispatchers stop finding readings that the record shows as current but that are actually old. For a route task, it might be that approved drafts need fewer manual edits over time. The criterion should be observable in the office's own data, not in a general impression.

How Be Human's managed approach handles this

Be Human designs and operates AI agents with their identity, memory, channels, tools, permissions and operating context kept under human control. A first engagement defines one useful role for a propane operation, such as tank records or route drafts, along with the memory and tools that role needs, the approval boundaries and a success criterion the office can observe. Safety notices stay on a human path from the start. Managed operations can then include monitoring, recurring evaluations, integration maintenance, permission reviews, incident response and controlled improvements. The service is remote and available in English and Portuguese.

If you want to find the first task worth handing to an agent, the free Be Human AI diagnostic is a practical place to begin.

Frequently asked questions

Can an AI agent decide when a propane tank needs a refill? It can project levels from the last reading and the consumption profile and flag tanks that fall below the office's threshold. The decision to schedule a delivery should still be approved by a dispatcher, because the route, the access and the driver's hours all affect whether a fill is practical. The agent's projection should always be shown as an estimate with its source.

Should customers receive refill confirmations from an AI agent? An agent can send confirmations written in approved wording once the dispatcher has placed the stop on an approved route. It should not confirm a window that is still a draft, and it should not answer questions about odor, leaks or damage. Those questions go to a person immediately.

What happens when a tank reading is missing or old? The record should show the tank as stale rather than current, so the dispatcher treats the level as unknown. The agent can place the tank in a separate group for a driver to read on the next visit. It should never fill the gap with a guess presented as a measurement.

How should a propane company begin without changing its safety procedures? Begin with one narrow record or route task and keep every safety-related notice on its existing human path. Test the escalation route before anything else goes live, and keep the approved script under the operations team's control. The existing procedures remain the authority, and the agent works inside them.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →